Target Detection Algorithm for UAV Images Based on Improved YOLOv8
Aiming at the problems of multiple small targets aggregation and large target scale variation in UAV aerial images,an improved YOLOv8 target detection algorithm named TS-YOLO(tiny and scale-YOLO)is proposed.Firstly,the redundant feature extraction layer is removed in the backbone part,and an efficient feature extraction module(EFEM)is designed to avoid small target features disappearing in redundant information.Secondly,a dual cross-scale weighted feature-fusion(DCWF)method is adopted in the neck,which fuses the multi-scale information and suppresses the noise interference for improving the feature expression ability.Finally,by constructing a parameter-shared detection header(PSDH),the regression and classification is took to realize parameter sharing,which ensures the detection accuracy and effectively reduces the number of parameters in the model.The precision(P)and recall(R)of the proposed model on the VisDrone-2019 dataset reach to 54.0％and 42.5％,respectively;compared with the original YOLOv8s model,the mAP50 is improved by 5.0 percentage points to 44.5％and the parameter quantity is reduced by 55.8％to only 4.94 ×106;on the DOTAv1.0 remote sensing dataset,the mAP50 reaches 71.9％,which still has good generalization ability.
万方数据
维普
